Microsoft plans to employ artificial intelligence to identify security flaws in the Windows operating system, with the aim of correcting them before cybercriminals can exploit them. The company introduced a tool called MDASH, which uses AI agents to search for vulnerabilities, positioning itself as the core element of this new defense strategy.
The use of AI aims to allow Windows security updates to contain a higher volume of fixes, without inherently needing to increase the frequency of these updates. The main goal is to accelerate the location of vulnerabilities in Windows, enabling their correction before attackers can take advantage of these gaps.
One of the mechanisms adopted by Microsoft in this process is the MDASH tool, which operates through a set of AI agents, characterized by the company as a 'multimodal agent scanning system.' This system was officially launched in May of this year and has since helped Microsoft detect 16 flaws, four of which were classified as 'critical.' All these vulnerabilities were patched in the same month they were identified.
Additionally, the company established a cloud-based infrastructure to support both the scans performed by MDASH and the validation of these findings. This structure's primary purpose is to prevent false positives from reaching the developers responsible for the fixes.
Microsoft strongly emphasizes that although the artificial intelligence resources serve to speed up the discovery and repair of vulnerabilities, all decisions regarding these procedures will remain under the purview of human teams. Furthermore, this new methodology will be applied not only to existing features but also to those in development, thus preventing the release of new functionalities or updates containing potential flaws.
The faster detection of flaws in Windows suggests that the respective corrections may be made available more quickly. However, this does not necessarily imply an increase in the periodicity of system updates, but rather that each update may cover a larger number of fixes. As mentioned by Pavan Davuluri, head of the Windows division, the advancement in the pace of vulnerability discovery ensures that customers no longer have to choose between speed and stability when implementing updates, as the focus is on keeping them securely protected.
Previously, cyberattacks were carried out by humans: hackers infiltrated systems, stole data, and demanded ransom. However, there is now a first case where the entire cyberattack was executed by an artificial intelligence (AI) agent.
Security researchers named this AI agent JadePuffer. They claim this is the first documented instance where AI independently performed almost all technical aspects of a ransomware attack. This news has caused serious concern in the cybersecurity field, as AI's ability to self-exploit could lead to faster and more dangerous attacks in the future.
Ransomware is a dangerous type of malware. When it enters a company's or private individual's system, it encrypts or locks files and data. Afterward, attackers demand a ransom for the return of the data; otherwise, the data may be permanently lost or published online. Until now, hackers made decisions at every stage of such attacks, but in the case of JadePuffer, the AI made many decisions autonomously.
According to the cloud security company Sysdig's research, JadePuffer first penetrated the server by exploiting an old vulnerability in the security system. It then independently scanned the system, found API keys and other passwords, gained access to other servers, and finally encrypted the data. Furthermore, the AI itself prepared the ransom note. Researchers noted a striking moment: when the AI encountered a problem at any stage, it corrected the error itself and continued the attack using the next method approximately 31 seconds later, eliminating the need for constant human instructions.
Although some initial reports suggested the entire attack proceeded without human involvement, Sysdig's Director of Research later clarified that a human selected the target, prepared the necessary infrastructure, and provided initial access. The technical part of the attack was then taken over by the AI agent. Thus, the AI did not act completely in isolation, but it performed many major tasks instead of the hacker.
Experts believe that if AI can independently penetrate systems, find data, search for passwords, correct errors, and demand ransom, such attacks could be conducted in large numbers and in a short time, making large-scale attacks easier even for small cybercriminals. Currently, the JadePuffer incident involves corporate servers, and there is no confirmation of attacks on personal phones or laptops of ordinary users. Nevertheless, if AI-based tools continue to improve, banks, hospitals, government institutions, and large corporations could become more frequent targets in the future, leading to negative consequences for ordinary citizens.
Cyber experts advise companies and institutions to update their software and servers promptly. It is critically important to immediately eliminate outdated vulnerabilities, as JadePuffer utilized a previously existing weakness. Additionally, the importance of using multi-factor authentication, regular backups, and continuous security monitoring has significantly increased.
The United States Cybersecurity and Infrastructure Security Agency (CISA) is employing the Mythos artificial intelligence model, created by Anthropic, to audit government software code, aiming to find security vulnerabilities. This news was revealed to Reuters by three sources familiar with the matter on Monday (6).
Sources indicated that this tool is being used to examine code repositories belonging to the U.S. government. The objective is to detect flaws that could be exploited by foreign espionage agents or cybercriminals. The Attack Surface Evaluation team, a unit of CISA responsible for penetration testing exercises and digital security assessments in government agencies, is conducting this analysis work.
Although two of the sources confirmed that the audits have already identified a considerable volume of vulnerabilities, they opted not to disclose specific details. Reuters did not obtain confirmation regarding the extent of the source code analyzed, nor the quantity, nature, or severity of the discovered flaws.
CISA's use of Mythos occurs amid tense relations between Anthropic and the U.S. government. The company, based in San Francisco and which made a confidential Initial Public Offering (IPO) request in the U.S., experienced a critical moment in February. During this period, it refused to remove protections in its artificial intelligence that would prevent its use in domestic surveillance systems or autonomous weapons.
Consequently, the U.S. Department of Defense formally classified Anthropic as a supply chain risk. This designation was previously applied only to foreign companies suspected of assisting in espionage activities. However, this determination was later blocked by a judge in March. Since then, the friction has decreased following the private launch of Mythos, an AI model described as extremely capable of identifying and exploiting cybersecurity breaches.
Previous reports published by Axios indicate that the U.S. National Security Agency (NSA) has been using Mythos since April, even while the risk classification was active. At the end of last month, The New York Times reported that NSA analysts were testing the model in confidential environments and were impressed by its capabilities. Subsequently, Anthropic released a public version of Mythos, named Fable, which, according to the company itself, contained cybersecurity-focused safeguards.
Following this release, the White House required Anthropic to restrict access to the model by international users. This order resulted in a global suspension of the tool, which was only lifted the previous week.
Reuters attempted to contact Anthropic, but the company did not provide responses regarding the initiative. A CISA representative stated last month that it would investigate the possibility of sharing data about the project but did not respond to subsequent contacts. Finally, the NSA declined to comment on the matter.
